wassname/TTS
1
0
Fork 0
mirror of https://github.com/wassname/TTS.git synced 2026-09-10 11:50:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

rename external speaker embedding arguments as d_vectors

Browse Source
This commit is contained in:
Eren Gölge
2021-06-28 17:03:19 +02:00
parent 9042ae9195
commit db6a97d1a2
20 changed files with 251 additions and 252 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_extract_tts_spectrograms.py
+3 -3
View File
@@ -22,7 +22,7 @@ class TestExtractTTSSpectrograms(unittest.TestCase):
c = load_config(config_path)
# create model
num_chars = len(phonemes if c.use_phonemes else symbols)
model = setup_model(num_chars, 1, c, speaker_embedding_dim=None)
model = setup_model(num_chars, 1, c, d_vector_dim=None)
# save model
torch.save({"model": model.state_dict()}, checkpoint_path)
# run test
@@ -41,7 +41,7 @@ class TestExtractTTSSpectrograms(unittest.TestCase):
c = load_config(config_path)
# create model
num_chars = len(phonemes if c.use_phonemes else symbols)
model = setup_model(num_chars, 1, c, speaker_embedding_dim=None)
model = setup_model(num_chars, 1, c, d_vector_dim=None)
# save model
torch.save({"model": model.state_dict()}, checkpoint_path)
# run test
@@ -60,7 +60,7 @@ class TestExtractTTSSpectrograms(unittest.TestCase):
c = load_config(config_path)
# create model
num_chars = len(phonemes if c.use_phonemes else symbols)
model = setup_model(num_chars, 1, c, speaker_embedding_dim=None)
model = setup_model(num_chars, 1, c, d_vector_dim=None)
# save model
torch.save({"model": model.state_dict()}, checkpoint_path)
# run test
tests/test_speaker_manager.py
+27 -27
View File
@@ -15,11 +15,11 @@ encoder_config_path = os.path.join(get_tests_input_path(), "test_speaker_encoder
encoder_model_path = os.path.join(get_tests_input_path(), "checkpoint_0.pth.tar")
sample_wav_path = os.path.join(get_tests_input_path(), "../data/ljspeech/wavs/LJ001-0001.wav")
sample_wav_path2 = os.path.join(get_tests_input_path(), "../data/ljspeech/wavs/LJ001-0002.wav")
x_vectors_file_path = os.path.join(get_tests_input_path(), "../data/dummy_speakers.json")
d_vectors_file_path = os.path.join(get_tests_input_path(), "../data/dummy_speakers.json")
class SpeakerManagerTest(unittest.TestCase):
"""Test SpeakerManager for loading embedding files and computing x_vectors from waveforms"""
"""Test SpeakerManager for loading embedding files and computing d_vectors from waveforms"""
@staticmethod
def test_speaker_embedding():
@@ -38,38 +38,38 @@ class SpeakerManagerTest(unittest.TestCase):
# load a sample audio and compute embedding
waveform = ap.load_wav(sample_wav_path)
mel = ap.melspectrogram(waveform)
x_vector = manager.compute_x_vector(mel.T)
assert x_vector.shape[1] == 256
d_vector = manager.compute_d_vector(mel.T)
assert d_vector.shape[1] == 256
# compute x_vector directly from an input file
x_vector = manager.compute_x_vector_from_clip(sample_wav_path)
x_vector2 = manager.compute_x_vector_from_clip(sample_wav_path)
x_vector = torch.FloatTensor(x_vector)
x_vector2 = torch.FloatTensor(x_vector2)
assert x_vector.shape[0] == 256
assert (x_vector - x_vector2).sum() == 0.0
# compute d_vector directly from an input file
d_vector = manager.compute_d_vector_from_clip(sample_wav_path)
d_vector2 = manager.compute_d_vector_from_clip(sample_wav_path)
d_vector = torch.FloatTensor(d_vector)
d_vector2 = torch.FloatTensor(d_vector2)
assert d_vector.shape[0] == 256
assert (d_vector - d_vector2).sum() == 0.0
# compute x_vector from a list of wav files.
x_vector3 = manager.compute_x_vector_from_clip([sample_wav_path, sample_wav_path2])
x_vector3 = torch.FloatTensor(x_vector3)
assert x_vector3.shape[0] == 256
assert (x_vector - x_vector3).sum() != 0.0
# compute d_vector from a list of wav files.
d_vector3 = manager.compute_d_vector_from_clip([sample_wav_path, sample_wav_path2])
d_vector3 = torch.FloatTensor(d_vector3)
assert d_vector3.shape[0] == 256
assert (d_vector - d_vector3).sum() != 0.0
# remove dummy model
os.remove(encoder_model_path)
@staticmethod
def test_speakers_file_processing():
manager = SpeakerManager(x_vectors_file_path=x_vectors_file_path)
manager = SpeakerManager(d_vectors_file_path=d_vectors_file_path)
print(manager.num_speakers)
print(manager.x_vector_dim)
print(manager.d_vector_dim)
print(manager.clip_ids)
x_vector = manager.get_x_vector_by_clip(manager.clip_ids[0])
assert len(x_vector) == 256
x_vectors = manager.get_x_vectors_by_speaker(manager.speaker_ids[0])
assert len(x_vectors[0]) == 256
x_vector1 = manager.get_mean_x_vector(manager.speaker_ids[0], num_samples=2, randomize=True)
assert len(x_vector1) == 256
x_vector2 = manager.get_mean_x_vector(manager.speaker_ids[0], num_samples=2, randomize=False)
assert len(x_vector2) == 256
assert np.sum(np.array(x_vector1) - np.array(x_vector2)) != 0
d_vector = manager.get_d_vector_by_clip(manager.clip_ids[0])
assert len(d_vector) == 256
d_vectors = manager.get_d_vectors_by_speaker(manager.speaker_ids[0])
assert len(d_vectors[0]) == 256
d_vector1 = manager.get_mean_d_vector(manager.speaker_ids[0], num_samples=2, randomize=True)
assert len(d_vector1) == 256
d_vector2 = manager.get_mean_d_vector(manager.speaker_ids[0], num_samples=2, randomize=False)
assert len(d_vector2) == 256
assert np.sum(np.array(d_vector1) - np.array(d_vector2)) != 0
tests/tts_tests/test_speedy_speech_layers.py
+2 -2
View File
@@ -57,7 +57,7 @@ def test_speedy_speech():
# with speaker embedding
model = SpeedySpeech(num_chars, out_channels=80, hidden_channels=128, num_speakers=10, c_in_channels=256).to(device)
model.forward(
x_dummy, x_lengths, y_lengths, durations, cond_input={"x_vectors": torch.randint(0, 10, (B,)).to(device)}
x_dummy, x_lengths, y_lengths, durations, cond_input={"d_vectors": torch.randint(0, 10, (B,)).to(device)}
)
o_de = outputs["model_outputs"]
attn = outputs["alignments"]
@@ -71,7 +71,7 @@ def test_speedy_speech():
model = SpeedySpeech(
num_chars, out_channels=80, hidden_channels=128, num_speakers=10, external_c=True, c_in_channels=256
).to(device)
model.forward(x_dummy, x_lengths, y_lengths, durations, cond_input={"x_vectors": torch.rand((B, 256)).to(device)})
model.forward(x_dummy, x_lengths, y_lengths, durations, cond_input={"d_vectors": torch.rand((B, 256)).to(device)})
o_de = outputs["model_outputs"]
attn = outputs["alignments"]
o_dr = outputs["durations_log"]
tests/tts_tests/test_tacotron2_model.py
+4 -4
View File
@@ -95,7 +95,7 @@ class MultiSpeakeTacotronTrainTest(unittest.TestCase):
criterion = MSELossMasked(seq_len_norm=False).to(device)
criterion_st = nn.BCEWithLogitsLoss().to(device)
model = Tacotron2(num_chars=24, r=c.r, num_speakers=5, speaker_embedding_dim=55).to(device)
model = Tacotron2(num_chars=24, r=c.r, num_speakers=5, d_vector_dim=55).to(device)
model.train()
model_ref = copy.deepcopy(model)
count = 0
@@ -105,7 +105,7 @@ class MultiSpeakeTacotronTrainTest(unittest.TestCase):
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for i in range(5):
outputs = model.forward(
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"x_vectors": speaker_ids}
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"d_vectors": speaker_ids}
)
assert torch.sigmoid(outputs["stop_tokens"]).data.max() <= 1.0
assert torch.sigmoid(outputs["stop_tokens"]).data.min() >= 0.0
@@ -259,7 +259,7 @@ class SCGSTMultiSpeakeTacotronTrainTest(unittest.TestCase):
stop_targets = (stop_targets.sum(2) > 0.0).unsqueeze(2).float().squeeze()
criterion = MSELossMasked(seq_len_norm=False).to(device)
criterion_st = nn.BCEWithLogitsLoss().to(device)
model = Tacotron2(num_chars=24, r=c.r, num_speakers=5, speaker_embedding_dim=55, use_gst=True, gst=c.gst).to(
model = Tacotron2(num_chars=24, r=c.r, num_speakers=5, d_vector_dim=55, use_gst=True, gst=c.gst).to(
device
)
model.train()
@@ -271,7 +271,7 @@ class SCGSTMultiSpeakeTacotronTrainTest(unittest.TestCase):
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for i in range(5):
outputs = model.forward(
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"x_vectors": speaker_embeddings}
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"d_vectors": speaker_embeddings}
)
assert torch.sigmoid(outputs["stop_tokens"]).data.max() <= 1.0
assert torch.sigmoid(outputs["stop_tokens"]).data.min() >= 0.0
tests/tts_tests/test_tacotron_model.py
+4 -4
View File
@@ -116,7 +116,7 @@ class MultiSpeakeTacotronTrainTest(unittest.TestCase):
decoder_output_dim=c.audio["num_mels"],
r=c.r,
memory_size=c.memory_size,
speaker_embedding_dim=55,
d_vector_dim=55,
).to(
device
) # FIXME: missing num_speakers parameter to Tacotron ctor
@@ -130,7 +130,7 @@ class MultiSpeakeTacotronTrainTest(unittest.TestCase):
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for _ in range(5):
outputs = model.forward(
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"x_vectors": speaker_embeddings}
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"d_vectors": speaker_embeddings}
)
optimizer.zero_grad()
loss = criterion(outputs["decoder_outputs"], mel_spec, mel_lengths)
@@ -305,7 +305,7 @@ class SCGSTMultiSpeakeTacotronTrainTest(unittest.TestCase):
gst=c.gst,
r=c.r,
memory_size=c.memory_size,
speaker_embedding_dim=55,
d_vector_dim=55,
).to(
device
) # FIXME: missing num_speakers parameter to Tacotron ctor
@@ -319,7 +319,7 @@ class SCGSTMultiSpeakeTacotronTrainTest(unittest.TestCase):
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for _ in range(5):
outputs = model.forward(
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"x_vectors": speaker_embeddings}
input_dummy, input_lengths, mel_spec, mel_lengths, cond_input={"d_vectors": speaker_embeddings}
)
optimizer.zero_grad()
loss = criterion(outputs["decoder_outputs"], mel_spec, mel_lengths)